Multicore computing

A chip off the old block
Multicore designs get round many of these problems. Using simpler, duplicated processor cores makes chips easier to design and build reliably, and enables performance improvements without excessive power demands.

In addition, multiple cores make finer-grained dynamic voltage scaling and power-down routines possible, again saving power.

Examining current research, the Berkeley team concluded that shallow six- to nine-stage instruction pipelines with in-order execution offer the best trade-off between performance and power consumption for general applications, and that this would allow a thousand cores on a single chip using an economical 30 nanometre (nm) fabrication process.

This may not be too far away either, as Intel has already produced processors using 45nm technology and Cisco has built a 188-core dedicated communications chip using 130nm fabrication that dissipates only 35W at 250MHz, while offering a throughput of 50 billion instructions per second.

Mix and match
So, is a chip with 1,000 identical processor cores really the most efficient use of silicon for maximum throughput? So far, multicore processors have simply duplicated two or four cores, with eight to come.

In the future, though, there might be additional benefits to be gained by paying attention to a theory formulated 40 years ago by computer mainframe pioneer Gene Amdahl.

He pointed out that all software has sequential operations that gain no benefit from parallel processing, and that adding a single larger processor to speed up the sequential code would produce higher throughput than adding 10 more small parallel cores.

For example, assume that 10 per cent of a program gets no extra speed on a 100-core processor. By replacing 10 of the cores with a single core dedicated to speeding up that 10 per cent of code, we end up with a much more efficient 91-core processor.

Such heterogeneous processors are still a research project rather than a reality. A single replicated processor core is much easier to design and build, and offers a standard platform for software. Trying to optimise multiple execution threads in a heterogeneous environment might be too difficult to manage efficiently.

Even so, Amdahl’s law may have some implications for operating system software developers at Microsoft and Apple in particular, as the number of cores grows beyond eight.

The other keys to manycore processing are the interconnections between the cores and the memory architecture that keeps them supplied with data.